When using MiniMax M2 model in OpenAI provider, the thinking block cannot distinguish between thinking content and body text. #9871